Upload README.md with huggingface_hub

README.md

@@ -32,7 +32,7 @@ One RTX 3080, one session, one fixed request (full methodology and CPU numbers i
 | concurrency 32 | Throughput | P50 |
 |---|---:|---:|
 | FastAPI + torch (cuda) | 37.8 rps | 844 ms |
-| Verdict (CUDA EP) | 408 rps | 78 ms |
+| Verdict (CUDA) | 408 rps | 78 ms |
 
 ## Files
 
@@ -50,7 +50,7 @@ Weights are unchanged from `convaiinnovations/laya`; this repo adds the ONNX exp
 - export parity: max logits difference 6.6e-6 against the official forward
 - token ids: id-for-id identical to the official pipeline
 - responses: round-4 probabilities identical (golden vectors, 40 cases per task)
-- serving on both execution providers, against a 2e-4 tolerance: CPU worst logits difference 1.1e-4, CUDA 5e-5; response layer bitwise on both
+- serving on CPU and CUDA, against a 2e-4 tolerance: CPU worst logits difference 1.1e-4, CUDA 5e-5; response layer bitwise on both
 
 ## License